How lakshyakushwahaaiml2023-ops/marginsense works
MarginSense is a patient-specific radiotherapy planning and research pipeline: it turns MRI/segmentation data into a physics-informed cell-density prediction, estimates ensemble uncertainty, and uses those estimates to support asymmetric treatment margins while protecting CSF/OAR regions. The graph also preserves conventional and GliODIL/PINN comparisons, evaluation, upload preprocessing, and visualization. Evidence is limited to the README and sampled source/index references; unsampled implementation links are omitted where their caller is not shown.
Main entry points: processed Processed patient volumes · amortized_pinn.py Amortized PINN · evaluate_pipeline.py Pipeline evaluation · evaluate_uncertainty.py Ensemble uncertainty
Components 23
Patient data
- MRI and labels
- Dataset preprocessing
src/preprocess.py - Upload preprocessing
src/preprocess_upload.py - Processed patient volumes
data/processed - Patient features
src/compute_features.py
Modeling
- Amortized PINN
src/models/amortized_pinn.py - Single-model training
src/training/train.py - Ensemble training
src/training/train_ensemble.py - Model checkpoints
outputs - Ensemble uncertainty
src/evaluate_uncertainty.py
Planning and evaluation
- Margin optimization
src/viz/server.py - Pipeline evaluation
src/evaluate_pipeline.py - Vanilla PINN baseline
src/baseline_vanilla_pinn.py - GliODIL baseline
src/baseline_gliodil.py - Uniform margin baseline
src/baseline_uniform_margin.py - Model comparison
src/compare_models.py - Predictions and reports
outputs
Upload and visualization
- Planning interface
src/viz/server.py - Visualization data export
src/viz/export_json.py - 3D mesh export
src/viz/export_meshes.py - 3D viewer
src/viz/index.html - Viewer data client
src/viz/data.js
Other components
- Clinician
Connections 35
- Clinician → Planning interface: submits patient
README.md - MRI and labels → Dataset preprocessing: preprocesses
README.md - MRI and labels → Upload preprocessing: uploads scans
src/preprocess_upload.py - Dataset preprocessing → Processed patient volumes: writes volumes
src/preprocess.py - Upload preprocessing → Processed patient volumes: writes bundle
src/preprocess_upload.py - Dataset preprocessing → Patient features: uses features
src/preprocess.py - Upload preprocessing → Patient features: uses features
src/preprocess_upload.py - Single-model training → Amortized PINN: trains
src/training/train.py - Ensemble training → Amortized PINN: trains
src/training/train_ensemble.py - Single-model training → Processed patient volumes: loads volumes
src/training/train.py - Ensemble training → Processed patient volumes: loads volumes
src/training/train_ensemble.py - Single-model training → Model checkpoints: saves weights
- Ensemble training → Model checkpoints: saves weights
- Amortized PINN → Patient features: uses covariates
src/models/amortized_pinn.py - Ensemble uncertainty → Amortized PINN: evaluates
src/evaluate_uncertainty.py - Ensemble uncertainty → Processed patient volumes: loads volumes
src/evaluate_uncertainty.py - Ensemble uncertainty → Model checkpoints: loads ensemble
src/evaluate_uncertainty.py - Ensemble uncertainty → Predictions and reports: writes uncertainty
src/evaluate_uncertainty.py - Margin optimization → Predictions and reports: reads predictions
src/viz/server.py - Margin optimization → Processed patient volumes: reads patient data
- Pipeline evaluation → Amortized PINN: evaluates model
src/evaluate_pipeline.py - Pipeline evaluation → Single-model training: runs training
src/evaluate_pipeline.py - Pipeline evaluation → Vanilla PINN baseline: runs baseline
src/evaluate_pipeline.py - Pipeline evaluation → GliODIL baseline: runs baseline
src/evaluate_pipeline.py - Pipeline evaluation → Predictions and reports: writes metrics
- Model comparison → Uniform margin baseline: compares margins
- Planning interface → Upload preprocessing: starts preprocessing
src/viz/server.py - Planning interface → Patient features: uses features
src/viz/server.py - Visualization data export → Amortized PINN: loads model
src/viz/export_json.py - Visualization data export → Processed patient volumes: reads volumes
src/viz/export_json.py - Visualization data export → Predictions and reports: reads predictions
src/viz/export_json.py - Visualization data export → Uniform margin baseline: uses margin baseline
src/viz/export_json.py - Planning interface → 3D viewer: serves interface
- 3D viewer → Viewer data client: uses viewer data
- 3D mesh export → 3D viewer: provides 3D assets